def train(create_model, X, y, batch_size, epochs, gpu_count, parameter_server,
method):
if gpu_count > 1:
ps_device = '/gpu:0' if parameter_server == 'gpu' else '/cpu:0'
with tf.device(ps_device):
serial_model = create_model()
if method == 'kuza55':
from keras_tf_multigpu.kuza55 import make_parallel
model = make_parallel(serial_model,
gpu_count=gpu_count,
ps_device=ps_device)
elif method == 'avolkov1':
from keras_tf_multigpu.avolkov1 import make_parallel, get_available_gpus
gpus_list = get_available_gpus(gpu_count)
model = make_parallel(serial_model,
gdev_list=gpus_list,
ps_device=ps_device)
elif method == 'fchollet':
# requires Keras (2.0.9?) https://github.com/fchollet/keras/commit/3dd3e8331677e68e7dec6ed4a1cbf16b7ef19f7f
from keras.utils import multi_gpu_model
model = multi_gpu_model(serial_model, gpus=gpu_count)
else:
model = serial_model = create_model()
print('Number of parameters:', serial_model.count_params())
model.compile(loss='categorical_crossentropy',
optimizer='sgd',
metrics=['accuracy'])
gauge = SamplesPerSec(batch_size)
model.fit(X, y, batch_size=batch_size, epochs=epochs, callbacks=[gauge])
gauge.print_results()
model.add(Flatten())
model.add(Dense(512))
model.add(Activation('relu'))
model.add(Dropout(0.5))
model.add(Dense(num_classes))
model.add(Activation('softmax'))
return model
with tf.device(ps_device):
serial_model = basic_model()
print('Serial model:')
serial_model.summary()
model = make_parallel(tower, gpu_count, ps_device)
print('Multi-GPU model:')
model.summary()
# initiate RMSprop optimizer
opt = keras.optimizers.rmsprop(lr=0.0001, decay=1e-6)
# Let's train the model using RMSprop
model.compile(loss='categorical_crossentropy',
optimizer=opt,
metrics=['accuracy'])
x_train = x_train.astype('float32')
x_test = x_test.astype('float32')
x_train /= 255
x_test /= 255
def cnn(channel=3):
"""
Architecture and model of the CNN.
:return: object, model of the CNN.
"""
sgd = optimizers.SGD(lr=0.01, momentum=0.0, decay=0.0, nesterov=True)
inputs = Input(shape=(40, 80, 1))
gaussian = TruncatedNormal(stddev=0.01, seed=None)
pitch = Conv2D(filters=32,
kernel_size=(32, 1),
kernel_initializer=gaussian,
activation='relu',
name='conv_1')(inputs)
pitch = BatchNormalization()(pitch)
pitch = MaxPooling2D(pool_size=(1, 80))(pitch)
pitch = Reshape((1, 9, -1))(pitch)
tempo = Conv2D(filters=32,
kernel_size=(1, 60),
kernel_initializer=gaussian,
activation='relu',
name='conv_2')(inputs)
tempo = BatchNormalization()(tempo)
tempo = MaxPooling2D(pool_size=(40, 1))(tempo)
bass = Conv2D(filters=32,
kernel_size=(13, 9),
kernel_initializer=gaussian,
activation='relu',
name='conv_3')(inputs)
bass = BatchNormalization()(bass)
bass = MaxPooling2D(pool_size=(4, 4))(bass)
bass = Reshape((1, 126, -1))(bass)
if channel == 2:
concatenate = Concatenate(axis=2)([pitch, tempo])
flatten = Flatten()(concatenate)
dense = Dense(200,
kernel_initializer=gaussian,
activation='relu',
name='dense_1')(flatten)
else: # channel == 3
concatenate = Concatenate(axis=2)([pitch, tempo, bass])
flatten = Flatten()(concatenate)
dense = Dense(400,
kernel_initializer=gaussian,
activation='relu',
name='dense_1')(flatten)
dropout = Dropout(0.5)(dense)
predictions = Dense(10,
kernel_initializer=gaussian,
activation='softmax',
name='dense_2')(dropout)
model = Model(inputs=inputs, outputs=predictions)
model = make_parallel(model, gpu_count=2)
model.compile(optimizer=sgd,
loss='categorical_crossentropy',
metrics=['accuracy'])
plot_model(model,
to_file='./plots/model_plot.png',
show_shapes=True,
show_layer_names=True)
return model
model = Model(inputs=input, outputs=output)
print('Single tower model:')
model.summary()
return model
tensorboard_dir = './tensorboard-logs/mnist_mlp_multi_ps_cpu/%d-gpu_%s' \
% (gpu_count, os.environ.get('CUDA_VISIBLE_DEVICES', ''))
with tf.device(ps_device):
serial_model = basic_model()
print('Serial model:')
serial_model.summary()
model = make_parallel(serial_model, gpu_count, ps_device)
print('Multi-GPU model:')
model.summary()
model.compile(loss='categorical_crossentropy',
optimizer=RMSprop(),
metrics=['accuracy'])
summary_writer = tf.summary.FileWriter(tensorboard_dir, sess.graph)
summary_writer.flush()
tensorboard_cb = keras.callbacks.TensorBoard(log_dir=tensorboard_dir)
history = model.fit(
x_train,
y_train,
batch_size=batch_size,
